Worked examples, and when to stop using them

Beginners and experts need opposite things

When you are new to a type of problem, studying a fully worked example teaches you more per minute than attempting the problem yourself. This is well established, and it is the one place where being shown the answer is the right move.

The reason is capacity. Working memory holds only a few items at once. A beginner attempting an unfamiliar problem spends all of it on searching (what do I do first, is this the right formula, what was that rule) and has nothing left for noticing the structure. A worked example removes the search, so attention goes to the pattern.

But this reverses. The same worked examples that help a beginner actively slow down someone who already knows the method: they now have to process an explanation of something they can already do, which costs capacity and gives nothing. This reversal has a name, the expertise reversal effect, and it is why study advice that ignores your current level is unreliable.

Fading, which is the actual technique

The right sequence is not examples or problems. It is a fade from one to the other:

1. Full worked example. Every step shown, with the reason for each.
2. Completion problem, last step removed. You do the final step.
3. Last two steps removed.
4. Only the first step given.
5. Nothing given.

Each stage is a small generation attempt at the point you are most ready for it. This structure, backward fading, because the removal starts from the end, outperforms both pure examples and pure problem solving in controlled comparisons.

Models are unusually good at generating this, and almost nobody asks them to:

Give me a worked example of this type of problem with the reasoning at every step. Then give me the same type of problem with the last step missing for me to complete. Then one with the last two missing. Then one with only the first step given. Then one with nothing. Wait for my answer at each stage before continuing.

That is a complete, well-evidenced practice sequence, generated in one message.

Study the example properly, which most people do not

Reading a worked example passively gets you very little. The version that works:

• After each line, cover the rest and predict the next step before revealing it.
• Write, beside each step, why that step is legal. Not what it does: why it is allowed.
• At the end, close the example and write the sequence of moves from memory, in words, without numbers. This is the pattern you are actually trying to acquire.

That last one is the difference between learning to do this problem and learning to recognise this kind of problem.

Knowing when to stop

The signal that you have crossed out of the worked-example stage is that you can predict the next step correctly before revealing it, twice in a row. From that point, examples cost you time. Move to completion problems immediately, and to blank problems soon after.

Students overstay in this stage badly, because reading examples is comfortable and pleasant and feels like work. The number of worked solutions somebody has read is close to uncorrelated with what they can do; a solution read is a solution you can follow.

The trap specific to AI

A model will produce an unlimited supply of worked examples on demand. That makes it easy to spend an entire evening in stage one, which feels productive and is barely study at all. The fix is procedural: decide in advance how many worked examples you will read before moving to completion problems. Two is usually right; three if the material is new. Then move, whether or not you feel ready, because feeling ready is not a reliable signal here.

Checking that the example is right

One caution the classic research did not have to deal with. Worked examples in a textbook were checked. A generated one was not. In any subject where the steps matter, verify the example against a source before you learn from it: a textbook problem, a past paper with a mark scheme, a symbolic tool like SymPy or Maxima for the algebra. Learning a wrong worked example is worse than learning nothing, because you will apply it confidently.

The safest arrangement: take the worked example from your course materials, and use the model to generate the fading sequence of practice problems around it.